Modal

We're at NVIDIA GTC Berlin this week 🇩🇪 Ollie Allman, Tom Davies, Benjamin Cowen, Carl Winkler and Ellen Lee are on the ground, so hit us up if you'd like to meet. Tomorrow (Oct 21), we're hosting something a bit different with Black Forest Labs, Factory and NVIDIA: live music and generative installations from the artists and creators defining the medium. Come hang out. RSVP link in the comments.

Solve a real biology problem using the latest advances in AI! On November 14th in Berkeley, we’re bringing biologists, AI engineers and software builders together for DNAForge’s AI × Biology Hackathon. → Tackle challenges in therapeutics, biosecurity or agricultural resilience with compute and scientific tools supported by our sponsors. → Meet companies that are hiring and get in-person feedback from scientists, engineers and founders. → Compete for $1,500+ in cash prizes, credits, and opportunities to commercialize your work. Come with teammates or on your own. Everyone applies individually, and we’ll help admitted participants form teams before the event. Applications take under 5 minutes. Apply by October 25th for priority consideration: https://lnkd.in/gkv8HJkr Supported by Modal, Boltz, Tamarind Bio, GXL, Talus Bio, BenchFlow, Synphony (YC P26), Biobase, Raycaster and Carbon Fellowship. Our judges and mentors include Ali Mottaghi (Stanford University), Alex Telford (Convoke), Sam Seymour (Biobase), Roya Amini-Naieni (Trilobio), Aqib Hasnain (Mithrl) and Ben Shababo (Modal), and many more! Co-hosted with Muslim Tech Collaborative at UC Berkeley

We recently got our hands on NVIDIA's new Vera CPU and put it to the test for agentic workloads on Modal Sandboxes. Running hundreds of concurrent Sandboxes on a single host gets tricky. Agent workloads are unpredictable, with long stretches of inactivity interrupted by sudden CPU spikes. Packing more workloads onto a machine improves throughput, but typically comes at the expense of latency as processes compete for resources. To see how Vera handles this, we built Agentless Agent Bench, a collection of 50 real-world agentic tasks covering compilation, file operations, git commands, and more. We tested Vera against x86 and Arm alternatives, measuring latency and throughput as we increased concurrency. Our early results show Vera delivering up to 2x the throughput of comparable x86 hardware, with remarkably little slowdown under heavy load. As more workloads pile up, other CPUs hit throughput plateaus, while Vera just keeps going. For Modal Sandboxes, this could mean packing more workloads onto each host, with higher throughput and more room to burst without sacrificing latency. We wrote about how we built the benchmark, what we found, and what we believe might be Vera's most interesting feature. Read the full post: https://lnkd.in/gbi7ijVT

One of the secrets to Modal’s speed is ImageFS, our custom filesystem that loads files on demand. When we introduced VM Sandboxes, we wanted to bring the same sub-second cold starts and fast performance to VMs. To make that possible, we built a new filesystem stack that rethinks how VMs access files. The result is dramatically faster filesystem operations, including 12x faster package installs, without sacrificing startup speed. We wrote about the engineering challenges, tradeoffs, and decisions that got us there: https://lnkd.in/g-YNfZuq

Reflection Beam is now available on Modal. Beam is a 501B-parameter open-weight MoE (23B active) from Reflection built for coding, agentic workflows, and reasoning. It supports a 1M token context window, with token efficiency a core focus. In early access, the Oura Health Intelligence team found that "Beam had fewer context leaks and was much more grounded in its answers than other frontier models we evaluated.”

Nemotron 3.5 Super VL from NVIDIA is live on Modal today with 120B params for visual reasoning and agentic work across images, docs, and UIs. Hear more about the thinking behind the Nemotron family: At Runtime, @Bryan Catanzaro explained why “efficiency is also intelligence,” and how NVIDIA pre-trained Nemotron 3 Ultra entirely in FP4.

Muse Spark Open, a coding-optimized open weight model from Meta, is now available on Modal. Muse Spark Open is the most capable US open weight model to date and is trained specifically for agentic coding and long-horizon tasks; it supports native multimodal understanding and supports a 1M token context window.